This PhD grant, funded by Fondazione Bruno Kessler, aims to develop an advanced Physics-Informed Machine Learning (PIML) framework for modeling complex hydrodynamic and environmental systems. By integrating physical principles with data-driven methods, the research will focus on optimizing next-generation irrigation strategies. The project will harness the synergy between physical laws (as implemented in the GEOSPACE system) and machine learning to enable predictive, real-time, and scalable modeling tools for sustainable water resource management.
Key Objectives
- Design hybrid PIML models that combine governing equations with data-driven predictive models (e.g., neural networks);
- Improve predictive accuracy and generalizability across heterogeneous environmental conditions;
- Incorporate real-time sensor data inputs to refine model states and parameters;
- Benchmark PIML approaches against traditional numerical solvers and/or black-box machine learning models.
Methodological Approach
The core innovation of this project lies in the integration of physical constraints (as derived from GEOSPACE) into machine learning models. Building on recent advances in PIML, the goal is to design, develop, and validate models that enforce conservation laws and boundary conditions within neural architectures. This may involve:
- Embedding partial differential equations (PDEs) directly into the loss functions of machine/deep learning models;
- Developing adaptive training strategies to trade-oO data fidelity and physical consistency;
- Utilizing sensor data to dynamically assimilate environmental variability into model predictions;
- Leveraging high-performance computing to train and deploy models at scale across complex
- domains.
Expected Outcomes
This integration will enable:
- Accurate and efficient modeling of water distribution and use in precision irrigation systems;
- Real-time monitoring and decision-support capabilities for agricultural and environmental applications;
- Enhanced data efficiency and model robustness through physics-based regularization;
- Improved understanding of system dynamics under data-scarce and/or non-stationary conditions.
Implementation Timeline
Months 1–6: Literature review on PIML methodologies;
Months 7–18: Design and develop a core PIML architecture, integrating IoT data sources;
Months 19–24: Validate models using lab-scale and field experimental datasets;
Months 25–36: Upscale models to real-world irrigation systems deployed within ongoing local and EU
projects (e.g., IRRITRE, AGRIF .OODTEF), and quantitatively assess their impact on water-saving strategies.
Possible Collaborations
Fabio Antonelli, Fondazione Bruno Kessler; Sara Bonetti and Concetta D'Amato, EPFL
Info: abouthydrology <at> gmail.com
